Within-season vegetation indices and yield stability as a predictor of spatial patterns of Maize (Zea mays L) yields

Accurate evaluation of crop performance and yield prediction at a sub-field scale is essential for achieving high yields while minimizing environmental impacts. Two important approaches for improving agronomic management and predicting future crop yields are the spatial stability of historic crop yields and in-season remote sensing imagery. However, the relative accuracies of these approaches have not been well characterized. In this study, we aim to first, assess the accuracies of yield stability and in-season remote sensing for predicting yield patterns at a sub-field resolution across multiple fields, second, investigate the optimal satellite image date for yield prediction, and third, relate bi-weekly changes in GCVI through the season to yield levels. We hypothesize that historical yield stability zones provide high accuracies in identifying yield patterns compared to within-season remote sensing images.

To conduct this evaluation, we utilized biweekly Planet images with visible and near-infrared bands from June through September (2018–2020), along with observed historical yield maps from 115 maize fields located in Indiana, Iowa, Michigan, and Minnesota, USA. We compared the yield stability zones (YSZ) with the in-season remote sensing data, specifically focusing on the green chlorophyll vegetative index (GCVI). Our analysis revealed that yield stability maps provided more accurate estimates of yield within both high stable (HS) and low stable (LS) yield zones within fields compared to any single-image in-season remote sensing model.

For the in-season remote sensing predictions, we used linear models for a single image date, as well as multi-linear and random forest models incorporating multiple image dates. Results indicated that the optimal image date for yield prediction varied between and within fields, highlighting the instability of this approach. However, the multi-image models, incorporating multiple image dates, showed improved prediction accuracy, achieving R² values of 0.66 and 0.86 by September 1st for the multi-linear and random forest models, respectively. Our analysis revealed that most low or high GCVI values of a pixel were consistent across the season (77%), with the greatest instability observed at the beginning and end of the growing season. Interestingly, the historical yield stability zones provided better predictions of yield compared to the bi-weekly dynamics of GCVI. The historically high-yielding areas started with low GCVI early in the season but caught up, while the low-yielding areas with high initial GCVI faltered.

In conclusion, the historical yield stability zones in the US Midwest demonstrated robust predictive capacity for in-field heterogeneity in stable zones. Multi-image models showed promise for assessing unstable zones during the season, but it is crucial to link these two approaches to fully capture both stable and unstable zones of crop yield. This study provides opportunities to achieve better precision management and yield prediction by integrating historical crop yields and remote sensing techniques.